False Positive Reduction in Lung Nodule Detection using Patch based Convolution Neural Networks

Abstract

Lung cancer is the leading disease among all the types of cancers in the world. Early prediction and diagnosis of lung cancer will increase the endurance rate of the affected patients. Therefore, in this paper, we proposed an efficient Computer-Aided Diagnosis (CADe) system for the automatic detection of pulmonary nodules and the reduction of false positives per scan from lung Computed Tomography (CT) images. Major steps in the design of the CADe system includes preprocessing of lung slice using Riesz filter banks, image segmentation using iterative thresholding approaches, nodule detection using morphological operators. The challenging task in the design of the CADe system is the reduction of false positives per scan. Here a novel patch-based Convolutional Neural Networks (CNN) are designed with three Convolution Layers with the number of filters as 32,64, and 128 respectively, two maximum pooling layers, Leaky ReLu activation function to increase the process of convergence during training. The final Fully-Connected (FC) layer produces two outputs as nodule and non-nodule. After detecting pulmonary nodules 64×64 patches are drawn, which are given as inputs to CNN. The proposed method is evaluated on the publicly available huge LIDC-IDRI database. Out of 1018 cases in our study, 888 scans containing 1186 nodules are considered. The results of our proposed CAD system show biased performance giving 94.8% sensitivity and 2.8 FP/scan compared to the state-of-the-art.